General-Purpose AI Is a Template. M&A Diligence Needs Precedent.

Anthropic just launched Claude Legal, a legal plugin for Claude. The legal tech world reacted like the sky was falling. It is not.

But when the company behind one of the most capable AI models in the world turns its attention to law, it matters. Claude Legal is part of a new generation of general-purpose AI tools for attorneys. These tools are all variations of the same idea: take a powerful general-purpose AI and apply it to legal work. That is genuinely exciting, and it is not the same thing as building software specifically for M&A due diligence.

The difference is one every M&A attorney already understands. A general contract template is not the same as your firm's battle-tested precedent.

A New Generation of Legal AI

Let's give credit where it is due. These tools are real, and they are impressive.

Claude Legal is an agentic framework built on top of Claude. You define playbooks and risk tolerances. It plans and executes multi-step legal tasks. It can connect to document management systems, Slack, Box, and other collaboration tools. It is fundamentally prompt-driven, but the prompts are structured as workflows rather than ad hoc questions.

Other tools in this category take similar approaches: fine-tuned LLMs with structured workflows for legal research and drafting, AI workspaces with chat-based analysis and tabular document review, or configurable platforms that layer domain-specific interfaces on top of general-purpose models.

What all of these tools share is more important than how they differ. They are built on general-purpose language models with a legal interface layered on top. All are configurable, flexible, and capable. All require the attorney to define what to look for, how to structure the output, and what constitutes a risk.

This is the template model. It is professional-grade. It is a real starting point. And it is fundamentally different from a precedent.

A Practical Law template covers the fundamentals. But your firm's precedent is a different category entirely. The precedent already knows which representations matter most, which indemnification structures your partners prefer, which risk allocations have been negotiated across dozens of prior deals. It encodes institutional knowledge that no template can replicate.

General-purpose legal AI is the template. Purpose-built M&A diligence software is the precedent. The knowledge is already in the product.

What a General-Purpose Tool Does Not Know About Your Data Room

When you point Claude Legal at a data room and define a playbook, you are the one supplying the M&A knowledge. The tool executes, but it does not know:

• That the assignment clause in a customer agreement is more important than the notice provision for M&A purposes
• That unlimited liability in an IP assignment warrants a red flag while the same language in a mutual NDA is standard market
• Which provisions to extract across 300 documents (you have to specify them)
• How to structure findings as a documents-by-provisions matrix a partner can scan in minutes
• How to cross-reference the purchase agreement against the data room to populate disclosure schedules

The same is true for every tool in this category. These tools are configurable, not preconfigured. The knowledge has to come from somewhere, and in every case, it comes from the attorney.

Consider a concrete example. Give any of these tools 50 customer agreements and ask them to find change-of-control provisions. They will find some. They will miss others where the language does not match their expectations. They will not automatically identify the standard form, compare every contract against it, and surface only the three that deviate. They will not structure findings by counterparty or organize them in a format you can hand to a partner for review. You get answers. What you need is analysis.

We wrote about this broader dynamic in The F1 Engine Problem: the most powerful AI engine in the world is not useful without the right chassis around it. In legal AI, that chassis is domain-specific infrastructure.

When M&A Expertise Is Engineered Into the System

A purpose-built platform does not wait for you to define a playbook. The playbook is the product.

Upload 300 documents. No prompts, no playbooks, no configuration. The system already knows what to do with a data room.

Automatic classification. Customer agreements, employment agreements, IP assignments, leases, NDAs, equity documents. The system categorizes every document because it was built to understand M&A document types, not because someone wrote a classification prompt.

Key provisions extracted across all documents simultaneously. Not because you asked for specific provisions, but because the system knows which provisions M&A attorneys need for each document type. Change-of-control, assignment, indemnification, limitation of liability, term, termination, non-compete, non-solicit. The extraction happens because the knowledge is engineered in.

Risk flagging with context. Red and yellow flags on provisions that warrant attorney attention, calibrated to M&A norms. An uncapped indemnity in an IP license gets flagged differently than one in a standard services agreement because the risk profile is different.

Structured output. A tabular grid, documents by provisions, that a partner can scan in minutes. Not a chat thread to scroll through. Not a playbook output to parse. A matrix built for the way deal teams actually review diligence.

Deal deliverables. Memos auto-generated from structured findings. Disclosure schedules populated by cross-referencing the purchase agreement against the data room. Every extraction linked back to the source text so attorneys verify findings rather than generate them.

This is not a better playbook. This is a fundamentally different product category. The M&A knowledge is not in your prompts. It is in the software. We explored this philosophy in The Pre-Move Thesis: the best AI does not wait for instructions. It anticipates what the attorney needs and does the work before anyone asks.

The Real Cost of "Close Enough"

The risk with general-purpose tools for M&A diligence is not wrong answers. It is unknown unknowns. The things the tool did not know to look for.

Variance detection. You have 50 customer agreements. Three deviate from the standard form in ways that create deal risk. A general-purpose tool analyzes each agreement in isolation, treating every document as if it were the first one it has ever seen. A purpose-built system identifies the standard form automatically and surfaces only the deviations, because that is what the attorney actually needs to review.

Disclosure schedule coverage. The purchase agreement requires disclosure of all contracts containing change-of-control provisions. No general-purpose playbook knows to cross-reference the data room against the merger agreement. A purpose-built system maps disclosure requirements to data room documents automatically.

Multi-model consensus. A single model extracts an indemnification cap from a 50-page agreement. Maybe it gets it right. Maybe it hallucinates a number from an adjacent clause. A purpose-built system runs multiple models in parallel and requires consensus before surfacing a finding, because accuracy in diligence is not a "pretty good" standard.

General AI Has Its Place. It Is Not the Data Room.

Claude Legal and other general-purpose legal AI tools are excellent for legal research, drafting, brainstorming, summarization, and ad hoc analysis. Use them for those tasks. They will make you faster.

M&A due diligence is a different kind of problem. It requires structured extraction across hundreds of documents, systematic risk flagging calibrated to deal context, institutional knowledge about which provisions matter for which document types, and auditable verification trails that let attorneys stand behind their work.

These are not features you can prompt your way into. They are the product of engineering M&A expertise directly into the system's architecture.

You would not use a Practical Law template to close a billion-dollar acquisition. You would use your firm's precedent. The same logic applies to AI.

---

The best AI models are commodities. The same models that power Claude Legal also power Mage. The differentiator is not the engine. It is what you build around it.

For M&A due diligence, what you build around it must encode decades of deal knowledge: which provisions to extract, how to flag risks, how to structure findings for deal teams, how to cross-reference documents against each other, how to verify results. Not as prompts an attorney writes on the fly, but as infrastructure that works the moment documents hit the data room.

That is not a playbook problem. It is a product engineering problem.

The next time someone asks why you need purpose-built diligence tooling when Claude just launched a legal plugin, ask them why your firm maintains precedent agreements when Practical Law exists.